2028 Matters More Than the 2027 Boom: Integrated Scenarios for Samsung Electronics & SK Hynix

Mon, 13 Jul 2026 20:39:17 +0900

Related context: This post is a follow-up to the HBM 2030 Supply-Demand Model Cross-Check, Samsung Electronics & SK Hynix Valuation Through 2028E Earnings, and the July 13 AI Hardware Selloff Analysis. Accepting 2027’s boom as the baseline, it quantifies the probability that supply, efficiency, and financing conditions shift simultaneously beginning in 2028.

TL;DR

HBM supply/demand is likely to remain tight through 2027. However, the projection of 2030 demand 26.7EB vs. supply 10.6EB, 2.52× deficit is not a base-case forecast — it requires multiple bullish assumptions to hold simultaneously.
The most important year is not 2027 but 2028. Supply expansion from Samsung Electronics, SK Hynix, and Micron, inference efficiency gains, and refinancing verification across AI data centers all converge in the same window.
HBM long-term supply contracts extend earnings duration but do not eliminate the cycle. They convert price risk into contract renewal risk, customer credit risk, product-mix risk, and repricing risk.
Conditional probabilities: P1 sustained excess demand 35%, P2 localized credit stress with hyperscaler re-concentration 40%, P3 efficiency gains and supply normalization 15%, P4 systemic credit seizure 10%.
Probability-weighted EPS for 2027: Samsung Electronics 60,750원, SK Hynix 393,000원. For 2028: Samsung 49,900원, SK Hynix 316,000원.
Present value of 2028 terminal values discounted at 11% annually: Samsung Electronics 317,246원, SK Hynix 1,956,316원 — implying 24.7% and 6.0% upside respectively to the July 13 KRX closing prices. Once 2028 normalization risk is incorporated, Samsung’s margin of safety is wider; SK Hynix is more sensitive to whether P1 holds.

Bottom Line

The 2027 boom is relatively visible. The harder question is whether high HBM pricing and earnings persist through 2028 and beyond. Current prices discount the duration of earnings well beyond the 2027 peak.

1. Analytical Questions and Evidence Quality

This analysis ties three questions together in a single framework.

How much confidence can we place in the view that HBM supply/demand will remain tight through 2027?
If supply, efficiency, and financing conditions change simultaneously in 2028, how do Samsung Electronics’ and SK Hynix’s earnings differ across scenarios?
Given those changes, at what implied earnings level does the current share price trade?

The numbers in this report must be read by category.

Category	Definition	Example in This Report
Fact	Verified from public filings or market data	July 13 KRX closing prices, Oracle FCF, CoreWeave borrowings
Consensus	Aggregated market estimates	2027–2028 EPS and net income
Broker estimate	Individual broker forecast	Korea Investment & Securities and BNK Investment & Securities SK Hynix EPS
Model output	Derived from disclosed formulas and assumptions	P1–P4 probability-weighted EPS and present values
Analyst judgment	Non-statistical analytical judgment	Scenario probabilities, fair PER, 11% required return
Blocked	Cannot be determined from public sources alone	HBM contract-level pricing, customer-by-customer volumes, actual production yield

Core formulas are as follows.

Scenario Terminal Value = Scenario EPS × Scenario Fair PER
Probability-Weighted EPS = Σ(Scenario Probability × Scenario EPS)
Probability-Weighted Terminal Value = Σ(Scenario Probability × Scenario Terminal Value)
Present Value = Terminal Value ÷ (1 + 11%)^Remaining Years
HBM Need/Fleet = Installed and active HBM demand stock ÷ supply capacity that can be served by installed fleet

Scenario probabilities are not derived from historical frequency distributions. They are conditional judgments reflecting currently observable demand, supply, efficiency, and financing conditions. Fair PER values are likewise not mechanically observed from the market but are analytical estimates incorporating earnings durability, business concentration, financial structure, and long-term risk.

2. Current Price and Market Consensus

July 13 KRX closing prices and consensus estimates aggregated on the same date are used throughout.

Item	Samsung Electronics	SK Hynix
2026-07-13 KRX Closing Price	254,500원	1,845,000원
2027 Consensus EPS	65,802원	444,359원
2028 Consensus EPS	65,907원	433,625원
2027 Consensus Net Income	44.3조원	32.4조원
2028 Consensus Net Income	44.1조원	31.9조원
Current Price / 2027 Consensus EPS	3.87×	4.15×
Current Price / 2028 Consensus EPS	3.86×	4.25×

On the surface, both companies look remarkably cheap. But these low PER multiples do not mean the market is unaware of 2027 earnings — they are more likely a signal that the market doubts the durability of earnings beyond 2028.

The dispersion among SK Hynix estimates illustrates this clearly.

Source	2027 EPS	2028 EPS	Interpretation
Market Consensus	444,359원	433,625원	Moderate 2028 normalization
Korea Investment & Securities	415,254원	495,696원	Earnings growth continues through 2028
BNK Investment & Securities	214,642원	66,989원	Rapid peak passage after 2027

For the same company, 2028 EPS estimates range from 66,989원 to 495,696원. The divergence stems not from 2026 demand forecasts but from differing assumptions on 2028 HBM average selling prices, the pace of supply expansion, and the duration of AI capex.

Recalculating current PER using this report’s probability-weighted EPS:

Year	Samsung Electronics	SK Hynix
2027	4.19×	4.69×
2028	5.10×	5.84×

Applying the probability-weighted fair PER in reverse to the current share price, the implied EPS the market is pricing in is approximately 31,700원 for Samsung and approximately 245,000원 for SK Hynix — roughly 52% and 45% below the respective 2027 consensus. The market is discounting not 2027 earnings per se, but how long those earnings will last.

3. What Has Changed in the Memory Cycle — and What Hasn’t

What Has Changed

AI servers, HBM, and enterprise SSDs have raised the share of enterprise and data-center demand in the total mix.
HBM requires customer qualification, advanced packaging, yield ramps, and long-term contracts, making short-run supply response slower than for commodity DRAM.
As HBM absorbs a growing share of wafer starts, commodity DRAM output from the same fabs declines, which can support commodity DRAM pricing.
Financing by large cloud providers and AI data center operators front-loads actual memory orders, acting as an accelerant.
Conversely, when credit problems emerge, demand does not fade gradually — it can step down abruptly through contract renegotiations and investment deferrals.

What Has Not Changed

High prices ultimately invite new fabs, yield improvement, process-node transitions, and competitor capacity additions.
High HBM pricing and margins increase the incentive for a second-source supplier to enter and for customers to apply cost-reduction pressure.
Long-term contracts do not eliminate price risk — they convert it into renewal pricing and customer credit risk.
Simultaneously applying peak EPS to a high PER is a double-optimism error that counts both earnings and multiples at their most favorable.

The claim that HBM has made memory a better business than legacy DRAM, and the claim that the memory cycle has been abolished, are not the same assertion.

4. HBM Supply-Demand Model and the 2028 Inflection

In the HBM 2030 Supply-Demand Model Cross-Check, we reproduced the original model’s demand 26.7EB, supply 10.6EB, 2.52× deficit in consistent units. The formula itself does not incorrectly compare a demand flow against a supply stock. The issue is that long-horizon assumptions multiply together in series.

These include:

Monthly token consumption growth
Model size, context length, and agent state expansion
KV cache and memory residency ratios
Throughput, quantization, and routing efficiency
HBM service life and migration of inference workloads to other memory types
Achievable production volumes and packaging yield across the three major memory makers

For this reason, the 2030 2.52× deficit is best treated as a bullish stress scenario rather than a base case. Re-deriving central assumptions aligned to each of P1–P4 gives the following:

Path	Token Multiplier	Model Scale	Throughput Efficiency	KV Efficiency	HBM Residency	Demand	Need/Fleet
P1 Orderly Excess Demand	20×	4.0×	12×	5.0×	65%	16.1EB	1.52×
P2 Localized Credit Stress & Re-concentration	18×	3.5×	12×	5.5×	60%	12.8EB	1.21×
P3 Efficiency Gains & Supply Expansion	12×	2.0×	14×	6.0×	50%	6.5EB	0.62×
P4 Systemic Credit Seizure	8×	2.0×	14×	7.0×	45%	2.6EB	0.25×

For comparability, supply capacity is held fixed at 10.6EB. Actual HBM wafer starts, product mix, good-die yield, and packaging capacity are non-public, so absolute values cannot be confirmed.

Looking along the time axis, the judgment simplifies considerably.

Period	Assessment	Confidence
2026–2027	HBM4 transition and AI demand outpace new-fab supply additions	High
2028	SK Hynix Yongin Line 1, Micron ID1/Tongluo, and Samsung HBM expansion begin easing supply	Medium-High
2029–2030	Remaining deficit is more likely to narrow than to deepen	Medium
2030 2.5× deficit	Requires multiple bullish assumptions to hold simultaneously	Low

The investment decision does not therefore hinge on whether the 2027 boom occurs. The pivot is whether the additional supply coming online in 2028 clears customer qualifications, and whether that timing coincides with inference efficiency gains and tighter AI data center financing.

5. SK Hynix Long-Term Contracts and the Earnings Curve

Korea Investment & Securities’ July 13 SK Hynix report attributes the Q2 consensus miss not to demand collapse but to a timing lag in how HBM product-mix shifts and long-term supply contract pricing flow through to reported results.

Item	2026F	2027F	2028F
Revenue	32.83조원	48.77조원	58.53조원
Operating Profit	24.51조원	37.45조원	44.70조원
EPS	292,068원	415,254원	495,696원
ROE	94.5%	65.4%	46.5%

HBM assumptions are also aggressive.

HBM Metric	2026F	2027F
Bit Growth	+23.0%	+31.2%
ASP	$1.60/Gb	$2.82/Gb
ASP Growth	-2.5%	+76.3%
Revenue	3.71조원	8.30조원
Revenue Growth	+28.0%	+123.6%

Key estimate adjustments within the report include:

HBM4 sales begin in Q3 2026, but the largest earnings inflection is modeled in Q4 when HBM ASP rises +52% QoQ.
Compared to the May estimates, non-HBM DRAM ASP for 2026–2027 was trimmed by approximately 16%.
2026 HBM revenue was raised by 17%.
2027 HBM shipment volume was held flat while ASP was cut by 5%.
NAND shipment volumes and pricing were raised.

This report is not a retraction of the long-term HBM growth thesis. It is a revision that lowers short-run commodity DRAM price elasticity and pushes back the point at which HBM4 economics enter reported results.

Why Long-Term Contracts Should Not Be Taken at Face Value

Contract-level price floors, ceilings, minimum purchase obligations, and prepayment terms are not publicly disclosed.
Reports that contracts carry “no price ceiling” cannot by themselves confirm that prices rise every year — one must know whether pricing is linked to spot rates and how often it is reset.
Long-term contracts lock in volumes and customer relationships but do not eliminate the risk of customer renegotiation, default, or credit deterioration.
Non-HBM DRAM and NAND outside those contracts remain fully exposed to the commodity price cycle.
If HBM4 yield is low, SK Hynix’s shipments and incremental earnings may still fall short even if the market remains undersupplied.

The largest benefit of long-term contracts is not cycle elimination but extending earnings duration and converting the form of risk.

6. How AI Infrastructure Financing Flows Into Memory Orders

AI memory demand is not driven purely by technology pull. Data center operators’ financing activity creates front-loaded orders, and when financial conditions tighten, order gaps can emerge sharply.

Rising AI Expectations
 -> Cloud/AI data center investment and long-term contract expansion
 -> Front-loaded capex in GPU, HBM, power, and data centers
 -> Asset, revenue, and enterprise value appreciation
 -> Additional debt and equity financing
 -> Larger capex programs

Reverse direction
Monetization delay or higher refinancing costs
 -> Contract renegotiation / expansion delays
 -> Memory order gaps
 -> ASP and share price decline
 -> Collateral value and financing conditions deteriorate
 -> Further capex reduction

The numbers confirmed from public disclosures are substantial.

Item	Confirmed Value	Interpretation
Oracle FY2026 Free Cash Flow	-$23.7B	Large-scale AI capex exceeds cash generation
Oracle Debt Financing	$43B	Heavy external funding for AI investment
Oracle Equity Financing	$5B	Debt alone is insufficient to cover investment
Oracle RPO	$638B	Long-term contracts provide high revenue visibility
Oracle Customer Prepayments & Direct GPU Provision	$75B	End customers share investment risk
CoreWeave Total Debt (End 2025)	$21.6B	High financial leverage in AI data centers
CoreWeave New DDTL 4.0	$8.5B	Additional financing to continue expansion
CoreWeave DSCR Threshold	Minimum 1.15×	Full refinancing verification begins no later than after June 30, 2027

CoreWeave’s end-2025 detailed borrowings comprised DDTL 2.0 approximately $5.04B, DDTL 2.1 approximately $2.74B, DDTL 1.0 approximately $1.55B, DDTL 3.0 approximately $0.34B, and equipment/software finance leases approximately $4.17B. DDTL 4.0, signed in March 2026, totals $8.5B.

These figures do not imply that AI demand is fictitious. The contracts, GPUs, power infrastructure, and data centers are real. However, if cash flow fails to keep pace with financing costs in 2027–2028, weaker operators are likely to sell assets to be absorbed by large cloud providers.

This is why the probability assigned to P2 — localized stress followed by asset re-concentration among hyperscalers — exceeds that of P4 systemic collapse. In P2, memory orders do not disappear but a gap of one or more quarters is plausible.

7. P1–P4 Conditional Probability Tree

Final probabilities are derived by sequencing real demand → financial verification → supply/efficiency → asset absorption, not from historical memory cycle frequencies.

2028 AI Monetization and Refinancing Verification
 Pass 50%
 ├─ P1 Sustained excess demand 70% = 35%
 └─ P3 Efficiency/supply normalization 30% = 15%
 Stress 50%
 ├─ P2 Localized stress/re-concentration 80% = 40%
 └─ P4 Systemic credit seizure 20% = 10%

The 50/50 first branch is not a statistical prior. It is a neutral starting point reflecting the fact that real demand and contracts are strong but financial verification is not yet complete. Calculating probabilities to single percentage-point precision from current data would produce false precision — adjustments are made only in 5-percentage-point increments.

Evidence Supporting Verification Pass vs. Financial Stress

Evidence Supporting Verification Pass	Evidence Supporting Financial Stress
HBM mass-production ramp timing lag through 2027	Oracle FY2026 FCF -$23.7B, debt financing $43B
Oracle RPO $638B, customer prepayments and direct GPU provision $75B	CoreWeave total debt $21.6B, new DDTL $8.5B
HBM long-term contracts and prepayments raise order visibility	CoreWeave DSCR 1.15× verification after mid-2027
Large cloud providers’ actual data center investment	Rising risk premium on long-dated AI bonds

The 70/30 split between P1 and P3 on the verification-pass branch reflects the fact that through 2027, qualification, yield, and packaging bottlenecks remain, while memory-side MoE, MLA, KV compression, routing, and offloading can all scale simultaneously from 2028 onward.

The 80/20 split between P2 and P4 on the stress branch is also clear. GPUs and data centers are not assets that disappear — hyperscalers with strong cash flow can acquire and redeploy them. Because simultaneous capex pullbacks, multiple covenant breaches, and HBM long-term contract impairments have not yet materialized together, P4 is treated as a tail risk at 10%.

Evidence Ledger

Evidence	P1	P2	P3	P4	Assessment
HBM supply lag through 2027	++	+	-	–	Shortage persists
Top-3 memory supply expansion post-2028	-	0	++	0	P3 independent pathway
CXMT client DDR5/LPDDR expansion	-	-	++	0	Caps commodity price ceiling before HBM
YMTC high-layer NAND ramp and customer adoption	-	-	++	0	Pressure on Samsung NAND and enterprise SSD mix first
China HBM/server memory customer qualification	–	-	++	0	If confirmed, shifts from P1 to P3
Oracle RPO & customer-provided GPU	+	+	0	-	Demand is real; buffers contagion
Oracle negative FCF & large financing	-	++	0	+	Localized credit stress risk
CoreWeave debt & DSCR verification	-	++	0	+	2027–2028 financial inflection
Rising AI long-dated bond risk premium	–	++	0	+	Shifts 5pp from P1 to P2
Inference efficiency gains & offloading	-	0	++	0	Still insufficient production-level evidence
Hyperscalers’ asset absorption capacity	0	++	0	–	Why P2 outweighs P4

++ and -- denote direction only and are not statistical likelihood ratios.

Probability Revision History

Step	P1	P2	P3	P4	Rationale
Initial	40%	35%	15%	10%	Physical supply shortage dominance, localized financial risk, supply/efficiency path
AI long-duration bond weakness reflected	35%	40%	15%	10%	P1 to P2 shift of 5pp; no contagion evidence
Hyperscaler capex review	35%	40%	15%	10%	Demand execution and asset concentration effects offset
AI long-term scenario review	35%	40%	15%	10%	2027–2028 unchanged; only post-2030 P3 probability expands
China supply path review	35%	40%	15%	10%	Commodity supply risk reflected; HBM customer qualification evidence lacking

Probability Uncertainty Ranges and Revision Rules

Path	Central	Uncertainty Range	Strong Signals That Would Shift the Central Estimate
P1	35%	25–45%	Two or more hyperscalers raise capex guidance; financing costs stabilize; long-term contracts strengthened
P2	40%	30–50%	Single weak-hand operator stress event with hyperscaler asset absorption
P3	15%	10–25%	Top-3 or China supply expansion confirmed in volume; per-token HBM consumption declines
P4	10%	5–15%	Multiple operator defaults, two or more capex cuts, and long-term contract impairments occur simultaneously

A single news article or anecdote warrants only a 0–2pp adjustment.
A single strong signal from a company filing or earnings report warrants at most a 5pp adjustment.
Two or more independent strong signals in the same direction warrant a 5–10pp shift.
Localized credit stress moves probability from P1 to P2; contagion evidence moves it from P2 to P4; supply/efficiency evidence moves it from P1 to P3.

8. Final Scenarios and HBM Price Pressure

Path	Prob.	2028 State	2030 Need/Fleet	HBM Price Pressure	Key Outcome
P1 Orderly Excess Demand	35%	Tight or balanced	1.3–1.8×	Flat to -10%	Contract price floor holds; HBM earnings plateau
P2 Localized Credit Stress & Re-concentration	40%	Order gap, modest supply slack	0.9–1.3×	-15 to -30%	Weak operators exit; hyperscalers absorb assets
P3 Efficiency Gains & Supply Expansion	15%	Growing supply surplus	0.6–0.9×	-20 to -35%	Token usage grows but HBM premium compresses
P4 Systemic Credit Seizure	10%	Capex cuts and inventory correction	0.3–0.6×	-40 to -55%	LT contract renegotiation, output cuts, capex cancellations, policy intervention

Holding all other variables constant, shifting 10 percentage points out of P2 changes the 2028 terminal values as follows:

Probability Shift	Samsung Terminal Value Change	SK Hynix Terminal Value Change	Meaning
10pp from P2 to P1	+21,800원	+227,000원	Higher probability of sustained undersupply
10pp from P2 to P3	-7,500원	-70,000원	AI economy grows but HBM premium compresses
10pp from P2 to P4	-16,800원	-132,000원	Localized stress contagion to systemic risk

SK Hynix is more sensitive to every probability shift. Even using the same central probabilities, SK Hynix’s valuation error bars and share price volatility are larger due to its higher HBM concentration.

9. China Supply Expansion Is a Variable Across All Paths, Not a Separate Scenario

CXMT and YMTC capacity growth is not a fifth scenario independent of P1–P4. It is a supply variable that alters product-level pricing and market share within every path.

China Progress Stage	Market Affected First	Samsung Electronics	SK Hynix	Scenario Mapping
CXMT domestic DDR4/DDR5/LPDDR expansion	Commodity DRAM	ASP and product mix pressure	Relatively limited direct impact	EPS trimmed across all paths; probabilities unchanged
CXMT global PC customer qualification	Global client DDR5/LPDDR	ASP ceiling and share pressure simultaneously	Client exposure at risk	P3 upside risk
China low-end server RDIMM penetration	Standard server DRAM	Partial server mix pressure	DRAM outside AI servers at risk	P3 upside review
YMTC high-layer NAND ramp and customer adoption	Consumer NAND/enterprise SSD	Largest NAND share and margin pressure	Solidigm and enterprise SSD at risk	P3 EPS and multiple reduction
China HBM mass production and AI customer qualification	HBM	Partial HBM catch-up offset	Large scarcity premium and share risk	5–10pp shift from P1 to P3
China ramp and global AI capex slowdown concurrent	All memory	Commodity/NAND/HBM combined pressure	HBM/DRAM combined pressure	Monitor for P3-to-P4 contagion

In 2026–2027, CXMT is more likely to cap the client DDR5 and LPDDR price ceiling than to directly undermine HBM pricing. YMTC may similarly intensify price competition in consumer NAND and some enterprise SSDs first.

For China to emerge as a direct HBM threat, DRAM die production alone is insufficient. TSV bonding, stacking, base-die integration, advanced packaging, and repeated mass-production qualifications by AI customers are all required.

The following signals must be confirmed before Chinese capacity expansion is incorporated numerically in meaningful quantities:

CXMT DDR5/LPDDR is adopted at volume by global PC makers at prices more than 20% below those of the top-3 memory suppliers.
CXMT secures qualification beyond consumer PCs into enterprise PCs and server RDIMM.
YMTC high-layer NAND is confirmed by actual shipment volumes and enterprise SSD customer wins — not just stated targets.
China HBM moves beyond sample qualification to repeated mass-production qualification by AI accelerator customers.
Supply growth and commodity DRAM/NAND price declines among the top-3 suppliers are observed together for two or more consecutive quarters.

10. Samsung Electronics: EPS and Fair Value by Scenario

Year	Scenario	EPS	Fair PER	Terminal Value
2027	P1	65,000원	8.5×	552,500원
2027	P2	62,000원	8.0×	496,000원
2027	P3	58,000원	7.5×	435,000원
2027	P4	45,000원	7.0×	315,000원
2028	P1	68,000원	8.5×	578,000원
2028	P2	45,000원	8.0×	360,000원
2028	P3	38,000원	7.5×	285,000원
2028	P4	24,000원	8.0×	192,000원

A base PER above 9× is not assigned to Samsung because foundry loss reduction has not yet been confirmed. An 8.5× P1 multiple becomes achievable only if HBM share recovery, commodity DRAM/NAND pricing, and foundry normalization all operate together.

An 8.0× multiple in P4 is applied to avoid the double-discount error of multiplying a trough EPS by an also-depressed PER. In an actual recession scenario, normalized EPS, net cash, and book value must all be considered together.

11. SK Hynix: EPS and Fair Value by Scenario

Year	Scenario	EPS	Fair PER	Terminal Value
2027	P1	430,000원	9.0×	3,870,000원
2027	P2	405,000원	7.0×	2,835,000원
2027	P3	370,000원	6.0×	2,220,000원
2027	P4	250,000원	5.5×	1,375,000원
2028	P1	470,000원	9.0×	4,230,000원
2028	P2	280,000원	7.0×	1,960,000원
2028	P3	210,000원	6.0×	1,260,000원
2028	P4	80,000원	8.0×	640,000원

The 9× P1 multiple presupposes HBM market share, long-term contract terms, high ROE, and a structural rerating of SK Hynix as a strategic AI asset. Maintaining this multiple through 2028 requires confirmed HBM4E share defense, long-term contract repricing, and shareholder return activity.

Korea Investment & Securities’ target price of 3,800,000원 applies a 6.0× PBR to 12-month forward BPS — above the historical band peak of 5.0× cited in the same report. A 3,800,000원 target therefore requires not just earnings delivery but a structural rerating of the memory business.

In P2 and P3, HBM concentration shifts from an advantage to a vulnerability. When ASP falls, earnings and multiples can compress simultaneously.

12. Probability-Weighted Results and Risk-Adjusted Comparison

Year	Company	Prob.-Weighted EPS	Blended Fair PER	Prob.-Weighted Terminal Value	PV at 11% Annual Discount	vs. Current Price
2027	Samsung Electronics	60,750원	8.04×	488,525원	421,385원	+65.6%
2027	SK Hynix	393,000원	7.53×	2,959,000원	2,552,333원	+38.3%
2028	Samsung Electronics	49,900원	8.18×	408,250원	317,246원	+24.7%
2028	SK Hynix	316,000원	7.97×	2,517,500원	1,956,316원	+6.0%

Four observations follow from these results:

Viewed through 2027 alone, both companies show risk-adjusted values above current prices.
Once 2028 normalization probability is discounted, Samsung’s margin of safety is wider.
SK Hynix’s 2028 present value is only 6% above the current share price. A large portion of expected return depends on P1 holding.
Samsung’s commodity memory, foundry, and net cash position provide a partial downside buffer in P2 and P3.

What is already well known to the market is HBM undersupply and higher memory pricing. What may be underpriced in two directions:

If Samsung’s HBM catch-up coincides with commodity DRAM/NAND price recovery, both earnings and multiples could improve together.
If SK Hynix’s long-term contracts and HBM4E defend price floors and share even after 2028 supply additions, the current normalization discount may be excessive.

Conversely, the market may prove correct if supply, efficiency, and credit conditions all improve simultaneously in 2028, cutting peak earnings short.

13. How to Think About AI Long-Term Demand Beyond 2030

AI 2040 Plan A presents a path from approximately 20 million H100e in 2026 to 60 billion H100e by 2034, with AI compute power consumption reaching 5TW by 2034. However, the document itself describes this as a policy scenario, not a best-estimate forecast, and notes that certain variables were adjusted alongside the narrative.

Long-term scenarios are therefore not fed directly into 2027–2028 EPS estimates.

Period	P1 Excess Demand	P2 Re-concentration	P3 Volume Growth / Price Normalization	P4 Failure	Status
2026–2028	35%	40%	15%	10%	Current analysis probabilities
Post-2030 conditional baseline	25%	35%	30%	10%	Long-term judgment; not a statistical probability

If AI and robotics raise the efficiency of chip design, fab construction, and production, total memory bit demand may grow while the scarcity premium on HBM specifically narrows. HBM market size expansion and long-term PER expansion must therefore be evaluated separately.

As long-term memory demand grows, the opportunity set expands not only for HBM but also for commodity DRAM, NAND, enterprise SSDs, controllers, CXL, power, cooling, and interconnects. This is a relative advantage for Samsung, with its broader product portfolio.

14. What Needs to Materialize for This Analysis to Hold

Samsung Electronics

HBM4 and HBM4E customer qualifications must convert into actual mass-production volumes and revenue share.
Commodity DRAM and NAND contract price recovery must offset HBM investment costs, bonuses, and provisions.
Foundry losses must narrow and begin making a meaningful contribution to 2027–2028 earnings.
Net cash and shareholder return capacity must be maintained despite heavy capex.

SK Hynix

HBM4E share and long-term contract price floors must hold even as Samsung and Micron expand supply.
HBM repricing beginning in Q4 2026 must track toward the 2027 average of $2.82/Gb.
TSV and stacking yield weakness must prove temporary and attributable to product transition costs.
High free cash flow must flow toward shareholder returns and capital policy, not just capacity expansion.

15. Risk Signals and Leading Indicators

Risk	Leading Indicator	Samsung Impact	SK Hynix Impact	Analysis Adjustment
HBM long-term contract renegotiation	Price floor, prepayment, and minimum purchase weakening	Moderate	Very large	P2/P4 up
Hyperscaler capex guidance cut	Two or more simultaneous guidance cuts	Large	Very large	P2/P4 up
AI data center refinancing event	CoreWeave covenant terms/financing cost, Oracle funding	Moderate	Large	P2 up; P4 up if contagion
Accelerated supply expansion	Samsung/Micron volume-capable HBM output rises	Positive for HBM catch-up; negative for pricing	Very negative	P3 up
Sharp inference efficiency improvement	Per-token HBM bits decline	Broad product portfolio provides partial buffer	Negative	HBM multiple lower
China commodity supply expansion	CXMT/YMTC pricing, share, global customer qualifications	More negative for commodity DRAM/NAND	Client DRAM and Solidigm under pressure	Product-level EPS reduction
China HBM challenge	TSV/stacking/base die/packaging/customer qualification	Partial offset of HBM catch-up value	Very negative for scarcity premium	P1 to P3 shift
Foundry normalization failure	Continued losses, no yield improvement	Samsung-specific risk	Limited impact	Samsung P1 multiple reduction

Upside Signals

2028 HBM long-term contracts with higher volumes and prices and broader customer diversification
Early stabilization of HBM4E yield and mass-production volumes
Simultaneous confirmation of Samsung’s rising HBM share and narrowing foundry losses
AI service revenue and cash flow growing faster than compute resource contracts
HBM lead times and allocations hold even as supply expansion lags
SK Hynix shareholder return plan becomes concrete

Downside Signals

HBM ASP decline and shipment volume drop occur simultaneously
Two or more major hyperscalers lower capex growth rates
Long-term contract renegotiation, breach, or prepayment impairment
Memory inventory build and spot-to-contract price inversion
Actual AI service per-token HBM consumption falls 20%+ for two consecutive quarters
Samsung/Micron volume supply increase coincides with HBM premium decline

On a monthly basis, the indicators to track together are: HBM4/4E contract prices and prepayments, top-3 memory producer volumes and yields, advanced packaging lead times, hyperscaler capex and cash flow, Oracle and CoreWeave financing costs, Taiwan server ODM revenue, and DRAM/NAND contract prices and inventory.

16. Relative Investment Interpretation

The two names are not equivalent ways to buy the same memory upcycle.

Scenario	Samsung Electronics	SK Hynix
P1 Strengthening	HBM catch-up and commodity memory tailwind	Largest earnings leverage from HBM share and LT contracts
P2 Warning	Business diversification and net cash provide partial defense	More sensitive to order gaps and multiple compression
P3 Strengthening	Broad memory product range and foundry execution provide partial buffer	Direct exposure to HBM premium compression
P4 Warning	Absolute decline unavoidable but relative defense possible	Greatest risk of simultaneous earnings and multiple decline

From a risk-adjusted standpoint, Samsung offers a wider and more defensive choice. SK Hynix has greater earnings leverage in P1, but its valuation shifts faster when supply, efficiency, and credit conditions change.

Equipment and materials names also should not be grouped together indiscriminately. Equipment companies exposed to order gaps must be separated from consumable and parts suppliers whose revenues track utilization rates. Equipment with confirmed purchase orders and components/materials with recurring revenue streams are a better fit for this scenario framework than a pure HBM theme basket.

17. Conclusion

The memory boom through 2027 can be maintained as the central path for earnings estimates. But the real test of current share prices is not 2027 EPS — it is the duration of earnings beyond 2028. The claim that the HBM supply deficit has lasted longer than past cycles, and the claim that memory companies deserve a permanently elevated scarcity multiple, are not the same argument.

Samsung Electronics holds HBM catch-up potential, commodity DRAM/NAND, foundry operations, and net cash simultaneously, offering P1 upside and relative downside defense across P2–P4. SK Hynix’s HBM share and long-term contracts provide the highest earnings leverage in P1, but if supply, efficiency, and financing conditions normalize, earnings and multiples can compress together.

The current probability-weighted results say: through 2027, both companies offer meaningful upside; looking through to 2028, Samsung’s risk-adjusted margin of safety is wider. For SK Hynix to see further rerating, confirmation is needed that HBM4E yield, 2028 long-term contract terms, and AI infrastructure financing continue to support P1.

Items That Cannot Yet Be Confirmed from Public Sources

Samsung Electronics’ segment-level EPS for 2027–2028 and foundry normalization sensitivity
SK Hynix long-term contract customer weighting, repricing formula, and prepayment/minimum purchase terms
Top-3 memory makers’ HBM-dedicated wafer input, good-die and stacking yield, and customer-specific producible volumes
Composition of 2028 hyperscaler capex between internal cash flow, corporate bonds, and project finance
CoreWeave and Oracle refinancing schedules and covenant headroom
Per-generation HBM4E capacity, bandwidth, and memory residency rate changes per accelerator
CXMT/YMTC actual shipment volumes, global customer qualifications, and product-level earnings sensitivity

Key Sources

The scenario probabilities and fair values in this report are analytical estimates based on publicly available information and do not constitute personalized investment advice. Actual investment decisions must separately account for price volatility, investment horizon, taxation, currency risk, and individual loss tolerance.

Scenario Analysis on Korea Invest Insights